VETT: VectorDB-Enabled Transfer-Learning for Time-Series Forecasting

Transfer learning is a technique that allows one to learn effective models for a given task, leveraging models that were previously trained on other tasks. In recent years, transfer learning has been applied in the context of time-series processing, with notable examples in anomaly detection and classification. However, a general, cross-domain, and scalable approach for transfer learning in time-series forecasting is missing. In this work we propose VETT, a novel approach that leverages feature similarity search (implemented through vector databases) and ensemble models for effective transfer learning for time-series forecasting. The architecture, designed to be deployed on the Cloud, is tested on benchmarks and state-of-the-art datasets to assess both its effectiveness and efficiency with respect to traditional solutions.